Detection of epileptiform spikes based on active learning

Epilepsy is a neurological disorder characterized by recurrent abnormal neuronal discharges. Electroencephalogram is often used clinically to assist in the diagnosis and treatment of epilepsy. The spikes and sharps contain a large amount of epilepsy-related pathological information, so the detection of spikes and sharps among the abnormal epileptic waveforms has more clinical diagnostic value. There are two problems in using machine learning to achieve automatic recognition of spike and sharp waves. One is that most of the EEG data are unlabeled data, and it is difficult to obtain a large number of labeled training sets; the other is that spikes and sharps are mixed with plenty of background waves, which lead to a data imbalance trouble. Based on the above backdrop, this paper implements a detection framework of epileptiform spikes using active learning in order to achieve better recognition results with as little cost as possible, and its major contributions are as follows: (1)The KNN attention layer is introduced in the learning engine to improve the generalization ability of the model in the case of few samples; (2)In terms of the sampling engine, MPGR (Manifold Preserving Graph Reduction) pre-processing is first performed to initially reduce the imbalance rate of the data and remove redundant points, then density-weighted uncertainty based on GAN is used to accelerate the efficiency of active learning, and finally boundary distance-based clustering sampling is used, which is to ensure diversity while taking balanced samples as much as possible. Results of experiments conducted on a hospital-supplied dataset show that the proposed framework is effective.